Langfuse vs LangSmith

By Camellia · Updated Nov 3, 2026

You've shipped an LLM feature, an agent, or a RAG pipeline, and now someone has asked the question that always arrives a month late: what did it actually do, and was it any good? You need traces to see what happened, evals to judge whether it was right, and a place to version the prompts so a quiet edit doesn't change production behaviour. Two products come up in nearly every shortlist for that job: Langfuse and LangSmith.

They overlap heavily on features. The real decision is about ownership and fit. Langfuse is open source at its core, framework-agnostic, and self-hostable by anyone, and as of January 2026 it belongs to ClickHouse. LangSmith is a closed-source platform from LangChain, tuned for LangChain and LangGraph apps, with self-hosting reserved for Enterprise contracts. This page is for the head of engineering or technical founder who has to pick one and defend the budget line. For the wider field, see the LLM observability roundup and the agent observability guide.

Key Facts

TL;DR

What Each Is Built For

Langfuse started as an open-source LLM engineering platform and stayed that way. Its README lists tracing, evaluations, prompt management, datasets and a playground as the core product, with SDKs for Python and JS/TS and integrations that run from the OpenAI SDK and LangChain to LlamaIndex, Vercel AI SDK, CrewAI and Haystack. The design bet is that your stack will change, so the observability layer shouldn't care which framework you picked this quarter.

LangSmith comes from the team that built LangChain and LangGraph. Its docs frame it around tracing production agents, building datasets from those traces, and running evaluations against them, and they list OpenAI, Anthropic, CrewAI, the Vercel AI SDK and Pydantic AI alongside the LangChain family. So it isn't locked to LangChain. But the closer your app sits to LangChain and LangGraph, the less instrumentation you write and the more of the product lights up by default.

Neither is a gateway or a model router. They sit beside your application and watch it. If you're still choosing the orchestration layer itself, LangChain vs LlamaIndex and the LangChain alternatives guide cover that decision first, because it narrows this one.

The ClickHouse Acquisition: What Changed and What Didn't

This is the 2026 development that most older comparisons miss. ClickHouse announced it had acquired Langfuse on January 16, 2026, alongside a funding round. Deal terms weren't disclosed.

Two things are worth keeping in perspective. First, Langfuse already ran on ClickHouse as its analytics store, so this is a company moving closer to its own foundation, not an unrelated buyer arriving. You can see it in the self-hosting architecture, where ClickHouse holds the observability data. Second, none of those commitments is a contract with you. If open-source continuity is a procurement requirement, pin your decision to the MIT licence on the core and keep a self-hosted exit path documented, which Langfuse's architecture makes realistic.

Tracing

Both platforms capture nested traces of LLM calls, tool calls, retrieval steps and agent actions, and both let you open a bad run and see where it went wrong.

The billing row matters more than it looks. Langfuse's billable units are the sum of traces, observations and scores, and units generated by features such as LLM-as-a-Judge, annotation queues and experiments count too. So a single user request that produces one trace with several LLM and tool observations burns several units, and every evaluator score adds one more. LangSmith bills on traces, so a deep agent run counts once at the trace level, but it has its own metering for other features.

Evaluations

This is where the products feel most similar, and where the differences are in workflow rather than capability.

The Langfuse queue limits come from its pricing page; the capability list comes from its README. On the LangSmith side, the evaluation concepts docs are more explicit about the offline/online split and about pairwise queues, which is a real point for teams that grade outputs by comparing two versions. If your eval practice is mostly "run a dataset against a prompt change and look at the diff," either tool does it. If you want to score live multi-turn conversations, LangSmith documents that path more directly.

Prompt Management

Both separate prompts from code so a non-engineer can edit and a release can pin a version. LangSmith's tag-to-commit model is a clean fit for promoting a prompt from staging to production without a deploy. Langfuse's caching story matters if prompt fetches sit on your hot path.

Self-Hosting and Licence

This is the cleanest difference, and for many teams the one that decides the match.

The practical read: Langfuse lets any team stand up the full tracing and eval stack inside its own network without a contract, and the self-hosting docs say which parts are team-maintained. LangSmith offers the same data-in-your-VPC outcome, but through an Enterprise sales process. If your traces can carry regulated or customer data that must not leave your environment, that gap decides it before pricing does.

Framework Fit

Don't over-read this table. LangChain apps trace fine in Langfuse, and non-LangChain apps trace fine in LangSmith. The difference is how much glue you write, and how much of your team's mental model already assumes one vendor's run tree.

Data Residency and Compliance

If you need Japan hosting, Langfuse is the only one of the two that lists it. If you need a vendor-run EU region, both do. If you need an on-premises deployment, see the section above.

Plan Tiers at a Glance

Langfuse Cloud

Source: Langfuse pricing. Core and Pro are monthly contracts paid by card; Enterprise needs a yearly commitment. Overage is graduated: $8.00 per 100k units from 100k to 1M, $7.00 from 1M to 10M, $6.50 from 10M to 50M, and $6.00 beyond 50M. Pro adds data retention management, higher ingestion limits and unlimited annotation queues, and the optional Teams add-on is $300/month.

LangSmith

Source: LangSmith pricing. After the included volume, traces are pay as you go, but the page doesn't print a per-1,000-trace rate, and upgrading base traces to extended retention carries an additional fee that isn't stated either. So LangSmith overage is Pricing on request as far as a public number goes.

Pricing at a Real Workload

Take a small production team: 5 engineers who need to see traces, and 1,000,000 Langfuse units a month. The Langfuse figure is derived straight from the published formula.

Two caveats keep this honest. A Langfuse unit isn't a LangSmith trace, so the same app produces fewer traces than units once observations and scores are counted, and the two columns aren't a like-for-like volume. And seats scale one way for each tool: LangSmith Plus grows with headcount, Langfuse Core and Pro don't. At 20 engineers, LangSmith Plus is 20 x $39 = $780/month before any trace overage, while Langfuse Core is still $29 plus its unit overage. The inverse also holds for a solo developer: LangSmith's free Developer seat and Langfuse's free Hobby tier are both enough to start.

For the team that is cost-sensitive at scale, the useful move is to count your own units. Run a week of production traffic on the free tier, read the unit total, and apply the graduated rates above.

Implementation and Time-to-Value

When Langfuse Is the Right Call

• You need the data to stay inside your own network and you don't want an Enterprise sales cycle to get there.
• Your stack mixes frameworks, or you expect to change orchestrators within a year.
• You want overage maths you can compute yourself from a public graduated rate.
• You have a larger team and a per-seat model would grow faster than your trace volume.
• You need Japan hosting, or open-source licensing is a procurement requirement.

When LangSmith Is the Right Call

• Your application is built on LangChain or LangGraph and you want the shortest path to rich traces.
• You evaluate live multi-turn conversations, or compare two outputs head to head with pairwise review.
• You're a small team where a handful of seats at $39 is simpler than counting units.
• You're an Enterprise buyer who wants a vendor-run hybrid deployment with an SLA and custom SSO and RBAC.
• You want prompt promotion through tags and commits to be a first-class workflow.

Decision Framework

If neither fits, the Langfuse alternatives guide and the Arize, Galileo and Braintrust comparison cover the other serious options. Start with one real service, instrument it through OpenTelemetry so you keep the option to switch, and revisit the choice when your trace volume or your compliance requirements change.